Search

Current fiber-optic distributed acoustic sensor (DAS) target recognition technologies continue to prioritize updating the feature learning tools while neglecting varying contributions of information from different dimensions. Few works focus on designing recognition algorithms from the perspective of multi-dimensional information extraction and utilization. In this paper, we introduce an end-to-end three-dimensional attention-assisted convolutional neural network (3-D ACNN) into DAS for the first time. This novel approach allows for automatic and concurrent extracttion of three-dimensional input in time, frequency, and space, aiming to cooperatively recognize sensing targets with higher accuracy. Our new scheme also includes comparative examination of 2-D and 3-D attention mechanisms to leverage the most efficient information in various dimensions, thereby enhancing recognition accuracy while maintaining computation efficiency. Field tests demonstrate that the 3-D CNN outperforms commonly used 2-D networks with only time-frequency (T-F), time-space (T-S), or space-frequency (S-F) inputs, improving recognition accuracy from 95.47% to 98.67%. Moreover, the proposed 3-D ACNN can further improve accuracy up to 99.33%, outperforming the basic 3-D CNN. It is also found that T-F information in DAS is richer and more recognized than T-S and S-F information. The use of 3-D attention yields better results than 2-D attention. Additionally, the processing time required for each spatial sensing point is only 0.22 ms on a commercial GPU (NVIDIA GeForce (R) GTX 1080 Ti), which is approximately 1/171 of the time required for the equivalent 2-D network with the same T-F input (38.52 ms). The 3-D ACNN enabled smart fiber-optic DAS (sDAS) represents a significant breakthrough, enabling fast and accurate multi-dimensional collaborative recognition. This advancement is expected to have a substantial impact for distributed sensing applications.

Fiber-optic distributed acoustic sensor (DAS) has been applied to various large-scale infrastructure monitoring areas in smart cities, leading to a new generation of fiber-optic Internet of Things for ground listening. However, its single-source detection and recognition methods may fail in unpredictable multi-source interfering environments in urban. When an unknown number of sources are nonlinearly mixed at the DAS's fiber receiver, it increases the difficulty of multiple source separation further. Therefore, in this article, it is proposed a novel multi-source separation method in fiber-optic DAS to separate individual vibration signals from the unidentified nonlinear mixing procedure with unknown number of sources. Firstly, the mixed source number is blindly estimated by utilizing the Gerschgorin disk estimator (GDE), which is effective and robust in real-field applications of DAS. Secondly, the statistically independent sources are separated with the extended slow feature analysis (xSFA) according to the nonlinear instantaneous mixing model constructed for DAS in this article, which considers the complexity of the vibration wave propagation to the subsurface fiber. It relies on the temporal correlation to recover structure of the source signals that has been destroyed in the nonlinear mixing procedure. Finally, evaluation indices for separation are studied and the effectiveness of both the multi-source separation and the source number estimation are verified through simulation experiments and field tests. Compared with the two benchmark methods of fast independent component analysis (FastICA) and the independent slow feature analysis (ISFA), it shows the complicated nonlinear mixture of DAS signals can be separated with higher reliability in both the artificially and the real-field mixed cases.